Pith. sign in

REVIEW 3 major objections 6 minor 48 references

ECLEIRS: Exact conservation law embedded identification of reduced states for parameterized partial differential equations from sparse and noisy data

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Writing solution and flux as row-wise divergences of a skew-symmetric field makes reduced-state predictions conserve exactly, even for parameters absent from training.

desk verdict Useful combination of divergence-free potentials and INR reduced states, but the exact-conservation claim applies to the modeled flux pair, not the physical flux; worth a serious referee after honest reframing. read the letter →

arxiv 2506.18855 v1 pith:L6NSO5E3 submitted 2025-06-23 physics.comp-ph physics.data-an

Aviral Prakash , Ben S. Southworth , Marc L. Klasky This is my paper · ORCID
classification physics.comp-phphysics.data-an
keywords exactconservationreducedstatedynamicsimplicitneuralrepresentationsparameterizedPDEssparseandnoisydataODEsshockpropagationskew-symmetricmatrixfield
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ECLEIRS is a reduced-state dynamics method for parameterized partial differential equations that builds the conservation law into the model architecture instead of adding it as a penalty. The paper's central claim is that because the modeled solution $q^m$ and flux $f^m$ are parameterized as the row-wise divergence of a skew-symmetric space-time matrix field, the conservation law $\partial q^m/\partial t + \nabla\cdot f^m = 0$ holds identically for every parameter, including parameters never seen during training. On 1-D advection, 1-D Burgers, and 2-D Euler shock problems, the authors report that ECLEIRS reconstructs clean signals from highly sparse, noisy measurements and forecasts unseen parameters more accurately than unconstrained latent-dynamics models or physics-informed penalty variants, while keeping the modeled conservation residual at machine precision. A sympathetic reader would care because experimental data is sparse and noisy, and multi-query tasks like parameter estimation and uncertainty quantification need cheap models whose predictions remain physically admissible outside the training set.

What carries the argument

The load-bearing object is the space-time divergence-free ansatz: for a $d$-dimensional spatial domain one introduces a skew-symmetric matrix field $A(w;\mu)\in\mathbb{R}^{(d+1)\times(d+1)}$ and sets $z^m=(q^m,f^m)^T$ to the row-wise divergence of $A$. Because row-wise divergences of skew-symmetric matrices have zero space-time divergence, the conservation law is embedded in the model form itself. The entries $a(w;\mu)$ are decoded by an autodecoder MLP from reduced states $\tilde q(\mu,t)=h_\theta(\mu,t)$, and the reduced-state dynamics are identified by fitting a neural ODE to automatic-differentiation derivatives. This design is what transfers exact conservation to the inference stage, as any predicted $\tilde q$ still maps into the space of divergence-free lifted fields. The paper notes the modeled flux $f^m$ is learned as an independent field and need not equal the physical flux $f(q^m)$; the exactly enforced law is $\partial_t q^m+\nabla\cdot f^m=0$.

What would settle it

Evaluate the residual $\partial q^m/\partial t+\nabla\cdot f(q^m)$ with the true constitutive flux (for example, $c q^m$ for advection or $\tfrac{1}{2}c(q^m)^2$ for Burgers) at validation parameters outside the training set; if this residual is far above machine precision while the modeled-flux residual is at machine precision, the exact conservation claim holds only for the learned flux pair, not for the physical conservation law.

Watch

Extended reading notes

Core claim

The discovery is that exact conservation and data-driven dimensionality reduction can be combined without any tunable penalty: take the lifted space-time vector $z^m=(q^m,f^m)^T$ and define it as the row-wise divergence of a skew-symmetric matrix field $A(w;\mu)$. The divergence of a row-wise divergence of a skew-symmetric matrix is identically zero, so $\partial q^m/\partial t + \nabla\cdot f^m=0$ is satisfied to machine precision for any space-time point and any system parameter, whether or not that parameter appeared in training. The entries of $A$ are generated by an implicit neural representation conditioned on latent states $\tilde q(\mu,t)$ produced by a hypernetwork, and the latent dynamics are learned as a parameterized neural ODE. The same construction guarantees local subdomain and global conservation of the integrated modeled solution by the divergence theorem. On the three shock-propagation test problems, the authors find ECLEIRS gives the most accurate predictions for unseen parameters under sparse and noisy data, with conservation errors around $10^{-8}$ in single precision.

Load-bearing premise

The guarantee is built for the learned flux, which can differ from the true physical flux; if the learned flux drifts away from the real one, conservation of the modeled pair does not imply the physics is conserved.

Editorial extensions

If this is right

  • Exact conservation becomes a hard guarantee of the reduced model, so forecasts for parameters or times outside the training window cannot drift out of the divergence-free constraint.
  • No penalty parameter needs tuning; the physics constraint is not a soft loss term whose strength depends on noise level or problem.
  • The mesh-free implicit representation lets the same reduced model ingest sparse sensor data and evaluate on any space-time locations, including different resolutions in offline and online stages.
  • The divergence theorem turns the pointwise constraint into local subdomain and global conservation of integrated quantities, which is the property many applications actually need.
  • The forecast stage inherits the guarantee because reduced states always map through the same divergence-free decoder after neural-ODE time integration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the machine-precision conservation is established for the modeled pair $(q^m,f^m)$; whether the physical flux $f(q^m)$ satisfies the law is a separate, empirical question, and the paper's conservation metric in Eq. (30) uses the physical flux, so this is checkable.
  • Editorial extension: because the potential formulation is linear, the method could be extended to coupled conservation systems by assigning each equation its own skew-symmetric potential, with a consistency penalty across shared variables as the paper suggests.
  • Editorial extension: the exact constraint may act as an implicit regularizer that explains much of the denoising gain; a clean ablation training the same architecture with the consistency loss but without divergence-free structure would isolate this effect.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript introduces ECLEIRS, a reduced/latent state dynamics method for parameterized PDEs that represents the modeled solution q^m and flux f^m as the row-wise divergence of a skew-symmetric matrix field A(w;μ) whose components are decoded from latent states via an implicit neural representation (Section 3.2, Eqs. (13)-(20)). This construction implies the modeled pair satisfies ∂q^m/∂t + ∇·f^m = 0 identically. The method is evaluated on 1-D advection, 1-D Burgers, and 2-D Euler problems for two tasks: recovering clean signals from sparse/noisy data and predicting dynamics at parameters unseen during training, with comparisons against unconstrained (IRS) and penalty-based physics-informed (PI-IRS) baselines across several sparsity and noise levels.

Significance. The idea of embedding a divergence-free ansatz into an INR-based latent dynamics model is a useful and timely contribution; the space-time divergence-free parametrization is clearly explained, and the benchmark study is extensive, covering three shock-propagation problems with varying sparsity, noise, interpolation, and extrapolation settings. The practical advantage that ECLEIRS avoids the penalty-parameter tuning required by PI-IRS is well supported by the experiments. However, the headline theoretical claim that the governing physical conservation law is satisfied exactly for unseen parameters is not established: the construction guarantees conservation only for the modeled flux pair, because f^m is an independent field with f^m ≠ f(q^m), as the manuscript itself states in Section 3.2. Once this distinction is made explicit, the empirical observation that physical residuals are O(1e-8) for advection and Burgers remains valuable, but it is a numerical finding rather than a structural guarantee.

major comments (3)
  1. [Section 3.2, Eqs. (13)-(22); abstract; Section 5.1.2] The exact-conservation guarantee applies to the modeled pair (q^m, f^m), not to the physical law ∂q^m/∂t + ∇·f(q^m) = 0. The manuscript states in Section 3.2 that 'as the flux is also modeled as f^m, therefore f^m ≠ f(q^m)', so the space-time divergence-free construction does not by itself enforce the governing conservation law. The abstract's claim that ECLEIRS 'satisfies conservation laws exactly even for parameters unseen' and the similar statements in Sections 5.1.2 and 5.2.2 should be restricted to the modeled flux, and the machine-precision physical residuals reported in Figures 7 and 15 should be explicitly presented as empirical results rather than as consequences of the architecture.
  2. [Section 5.3, Eq. (35)] For the 2-D Euler problem, only the continuity equation ∂ρ/∂t + ∇·(ρu) = 0 is embedded in ECLEIRS and PI-IRS; momentum and energy conservation are not enforced. The manuscript should state this limitation wherever global statements such as 'satisfies conservation laws exactly' appear, and the conclusion should not imply that all conservation laws of the Euler system are satisfied.
  3. [Section 3.2 and Section 6] The inconsistency between the modeled flux and the physical flux is acknowledged in Section 3.2 and deferred to future work in Section 6, but it is central to interpreting the method's guarantees. Because the exactness property holds for a different flux, the paper should report the consistency error ||f^m - f(q^m)|| on the training and validation sets, or otherwise quantify how close the learned flux is to the physical flux. Without such a quantification, the small physical residuals in the advection and Burgers experiments are difficult to interpret and the theoretical contribution remains incomplete.
minor comments (6)
  1. [Figure 1 and Figure 19 captions] The captions refer to panel '(e)100% sparsity', but only panels (a)-(c) appear in both figures; please correct the panel references.
  2. [Table 4] The third parameter column is labeled 'xin', but the Burgers problem uses ω as the third parameter; please fix the header.
  3. [Section 5.2.1] The phrase 'an anamalous behavior' contains a typo; it should read 'an anomalous behavior'.
  4. [Figure 20 caption] The caption contains the typo 'groundn truth'; it should read 'ground truth'.
  5. [Equation (22)] The equation appears with garbled repeated arguments 'qqqm(t,x;µt,x;µt,x;µ)'; please fix the typesetting so that the argument is written consistently as (t,x;μ).
  6. [Section 5.1.1] The description of ECLEIRS as 'parameter-free' is only true with respect to the penalty parameter λ; the method still depends on architecture sizes and latent dimension. Please qualify the wording to avoid overclaiming.

Circularity Check

1 steps flagged · score 4.0 of 10

Exact conservation is a definitional property of the modeled flux pair, not of the physical flux; the abstract's 'governing conservation law' claim conflates the two.

  1. self definitional [Section 3.2 (Eqs. 13 and 16) and Section 3.4 (Eq. 22); abstract]
    "Through algebraic manipulation, it is observed that these satisfy ∇w·zzz(www;µµµ) = 0. ... As the flux is also modeled as ffffm, therefore ffffm̸=fff(qmqmqm). As a result, there is an inconsistency between the true flux and the modeled flux. ... As the conservation laws are exactly satisfied, the conservation law for the modeled equation can be represented as ∂qqqm(t,x;µt,x;µt,x;µ)/∂t +∇·fffm(t,x;µt,x;µt,x;µ) =O(ϵ)."

    The exact-conservation guarantee is not a derived prediction but an identity built into the ansatz: z^m is defined as the row-wise divergence of a skew-symmetric matrix field, so ∇·z^m = 0 holds for every parameter, seen or unseen, by construction. The paper explicitly states f^m ≠ f(q^m), so the law guaranteed exactly is ∂q^m/∂t + ∇·f^m = 0, not the governing physical law ∂q^m/∂t + ∇·f(q^m) = 0. The abstract's claim that ECLEIRS 'satisfies the governing conservation law up to machine precision' is therefore circular if 'flux' means the modeled f^m: the property is true by definition of the representation. The independent, non-circular content is the empirical physical-flux residual in Eqs. (30) and (33), which is not implied by the construction and must be verified numerically.

full rationale

The paper's central conservation guarantee reduces, by its own equations, to the defining property of the skew-symmetric-potential ansatz: z^m is parameterized as the row-wise divergence of a skew-symmetric matrix, making ∇·z^m=0 an algebraic identity rather than a predicted outcome. Because Section 3.2 acknowledges f^m ≠ f(q^m), the exactly conserved law is not the physical conservation law unless the modeled flux happens to match the physical flux. The abstract's phrasing that ECLEIRS 'satisfies the governing conservation law up to machine precision' thus conflates a definitional property of the modeled pair with a physical-law guarantee; the genuinely empirical claim, supported by the physical-flux conservation errors in Eqs. (30) and (33), is that the learned q^m and f^m approximate the true q and f well enough for unseen parameters. This is a partial, limited circularity: the headline guarantee is true by construction for the modeled pair, while the physical conservation results retain independent empirical content. There is no load-bearing self-citation chain: the divergence-free representation is attributed to external prior work ([28], [29], [30]), and the unseen-parameter dynamics comparisons against IRS and PI-IRS are genuine empirical benchmarks, not fitted inputs renamed as predictions. Overall score 4 reflects one central claim that partially reduces to its own definition, while the method's broader contribution is not circular.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a standard vector calculus identity (treated as an axiom), the smoothness assumption of the MLP representation, and the ad hoc decision to model flux independently. The independent flux field is the most fragile invented entity because it determines what conservation means in the claim.

free parameters (2)
  • Latent state dimension n_r = n_r = d_mu + 2 (advection/Burgers: 5, 2-D Euler: 4)
    Chosen by hand as the intrinsic manifold dimension plus one, not derived from the PDE or data. Affects expressivity and conservation error.
  • Neural network architecture sizes = Varies per problem (e.g., 4 layers, 64 neurons for hypernetwork in 1-D)
    Set manually and kept fixed across methods to ensure fair comparison, but not shown to be optimal or necessary for exact conservation.
assumptions (4)
  • standard math Row-wise divergence of a skew-symmetric matrix field is divergence-free, i.e., ∇_w · (row-wise divergence of A) = 0 for smooth A.
    This is the vector calculus identity from [29,30,28] that underlies Eqs. (12)-(20). It holds under equality of mixed partials, which requires smoothness.
  • domain assumption The solution and flux fields are smooth enough for the MLP representation to approximate the true shock solutions.
    The divergence-free representation with a smooth MLP describes smooth divergence-free fields; shocks are captured only approximately, as seen in the overshoots in the Burgers experiments.
  • ad hoc to paper The modeled flux f^m can be learned independently of the solution q^m and the resulting pair (q^m, f^m) represents the physics correctly.
    Section 3.2 explicitly states f^m ≠ f(q^m). The physical meaning of the conservation law depends on the flux being a function of the solution, which is not enforced.
  • domain assumption The hypernetwork and decoder have sufficient capacity to represent the parameterized solution manifold for all parameters in the training and validation sets.
    The method relies on the universal approximation capability of MLPs; the paper uses fixed architectures and minimal tuning, so poor representation would degrade results.
invented entities (1)
  • Independent modeled flux field f^m(www;µ) decoupled from the modeled solution q^m
    purpose: Adds degrees of freedom so that (q^m, f^m) can satisfy the space-time divergence-free condition exactly, which is the core of ECLEIRS.
    The physical flux is a function of q, but here it is a separate output of the neural network. Section 3.2 acknowledges f^m ≠ f(q^m), so the conserved flux is a modeling construct rather than the true physical flux.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECLEIRS: Exact conservation law embedded identification of reduced states for parameterized partial differential equations from sparse and noisy data." pith.science (2026). https://pith.science/paper/L6NSO5E3

@misc{pith2026250618855,
  author       = {Pith},
  title        = {Pith review of: ECLEIRS: Exact conservation law embedded identification of reduced states for parameterized partial differential equations from sparse and noisy data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L6NSO5E3}},
  note         = {Machine review of arXiv:2506.18855}
}
read the original abstract

Multi-query applications such as parameter estimation, uncertainty quantification and design optimization for parameterized PDE systems are expensive due to the high computational cost of high-fidelity simulations. Reduced/Latent state dynamics approaches for parameterized PDEs offer a viable method where high-fidelity data and machine learning techniques are used to reduce the system's dimensionality and estimate the dynamics of low-dimensional reduced states. These reduced state dynamics approaches rely on high-quality data and struggle with highly sparse spatiotemporal noisy measurements typically obtained from experiments. Furthermore, there is no guarantee that these models satisfy governing physical conservation laws, especially for parameters that are not a part of the model learning process. In this article, we propose a reduced state dynamics approach, which we refer to as ECLEIRS, that satisfies conservation laws exactly even for parameters unseen in the model training process. ECLEIRS is demonstrated for two applications: 1) obtaining clean solution signals from sparse and noisy measurements of parametric systems, and 2) predicting dynamics for unseen system parameters. We compare ECLEIRS with other reduced state dynamics approaches, those that do not enforce any physical constraints and those with physics-informed loss functions, for three shock-propagation problems: 1-D advection, 1-D Burgers and 2-D Euler equations. The numerical experiments conducted in this study demonstrate that ECLEIRS provides the most accurate prediction of dynamics for unseen parameters even in the presence of highly sparse and noisy data. We also demonstrate that ECLEIRS yields solutions and fluxes that satisfy the governing conservation law up to machine precision for unseen parameters, while the other methods yield much higher errors and do not satisfy conservation laws.

Figures

Figures reproduced from arXiv: 2506.18855 by the authors.

Figure 1
Figure 1. 1-D advection problem: Space-time points for (a) [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. 1-D advection problem: (a) Solution and (b) error for [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. 1-D advection problem: Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (23 more)
Figure 4
Figure 4. Figure 4: 1-D advection problem: Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: 1-D advection problem: Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: 1-D advection problem: Relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: 1-D advection problem: Mean conservation error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: 1-D Advection problem: Box plots of relative errors (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: 1-D Advection problem: Box plots of (a) relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: 1-D Burgers problem: (a) Predicted solution and (b) error for [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: 1-D Burgers problem: Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: 1-D Burgers problem: Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: 1-D Burgers problem: Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: 1-D Burgers problem: Relative error in space-time (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: 1-D Burgers problem: Mean conservation error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p028_15.png]
Figure 16
Figure 16. Figure 16: 1-D Burgers problem: Box plots for relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p029_16.png]
Figure 17
Figure 17. Figure 17: 1-D Burgers problem: Box plots of (a) relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p029_17.png]
Figure 18
Figure 18. Figure 18: The selection of parameter points used in the learning dataset (blue markers) and [PITH_FULL_IMAGE:figures/full_fig_p031_18.png]
Figure 19
Figure 19. Figure 19: 2-D Euler problem: Spatial locations of data used to learn different reduced state [PITH_FULL_IMAGE:figures/full_fig_p032_19.png]
Figure 20
Figure 20. Figure 20: 2-D Euler problem: Density at (a) training data based on [PITH_FULL_IMAGE:figures/full_fig_p033_20.png]
Figure 21
Figure 21. Figure 21: 2-D Euler problem: (a) Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p034_21.png]
Figure 22
Figure 22. Figure 22: 2-D Euler problem: (a) Box plots for relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p034_22.png]
Figure 23
Figure 23. Figure 23: 2-D Euler problem: Predicted density at µ = [0.34, 0.63] (in the validation dataset) for (a) ground truth, (b) IRS, (c) PI-IRS with λ = 10−1 and (d) ECLEIRS. These reduced state dynamics models were trained using data from learning dataset with 20% spatial sparsity, 2…
Figure 24
Figure 24. Figure 24: 2-D Euler problem: Predicted density at µ = [0.527, 0.775] (in the validation dataset) for (a) ground truth, (b) IRS, (c) PI-IRS with λ = 10−1 and (d) ECLEIRS. These reduced state dynamics models were trained using data from learning dataset with 1% spatial sparsity, …
Figure 25
Figure 25. Figure 25: 2-D Euler problem: Box plots comparing relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p036_25.png]
Figure 26
Figure 26. Figure 26: 2-D Euler problem: (a) Box plots of relative error (defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p036_26.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 38 canonical work pages

  1. [1]

    Turbulence and the dynamics of coherent structures part III: Dy- namics and scaling.Quarterly of Applied Mathematics, 45(3):583–590, 1987

    Lawrence Sirovich. Turbulence and the dynamics of coherent structures part III: Dy- namics and scaling.Quarterly of Applied Mathematics, 45(3):583–590, 1987

    work page 1987

  2. [2]

    Lumley, and Emily Stone

    Nadine Aubry, Philip Holmes, John L. Lumley, and Emily Stone. The dynamics of coherent structures in the wall region of a turbulent boundary layer.Journal of Fluid Mechanics, 192:115–173, 1988

    work page 1988

  3. [3]

    A survey of projection-based model reduction methods for parametric dynamical systems.SIAM Review, 57(4):483–531, 2015

    Peter Benner, Serkan Gugercin, and Karen Willcox. A survey of projection-based model reduction methods for parametric dynamical systems.SIAM Review, 57(4):483–531, 2015. 37

    work page 2015

  4. [4]

    Aviral Prakash and Yongjie J. Zhang. Projection-based reduced order modeling and data-driven artificial viscosity closures for incompressible fluid flows.Computer Methods in Applied Mechanics and Engineering, 425:116930, 2024

    work page 2024

  5. [5]

    Benjamin Peherstorfer and Karen E. Willcox. Data-driven operator inference for non- intrusive projection-based model reduction.Computer Methods in Applied Mechanics and Engineering, 306:196–215, 2016

    work page 2016

  6. [6]

    Adjacency-based, non-intrusive model reduction for vortex-induced vibrations.Computers & Fluids, 275:106248, 2024

    Leonidas Gkimisis, Thomas Richter, and Peter Benner. Adjacency-based, non-intrusive model reduction for vortex-induced vibrations.Computers & Fluids, 275:106248, 2024

    work page 2024

  7. [7]

    Alberto Padovan, Blaine Vollmer, and Daniel J. Bodony. Data-driven model reduc- tion via non-intrusive optimization of projection operators and reduced-order dynamics. SIAM Journal on Applied Dynamical Systems, 23(4):3052–3076, 2024

    work page 2024

  8. [8]

    Kara, and Yongjie J

    Vedant Puri, Aviral Prakash, Levent B. Kara, and Yongjie J. Zhang. SNF-ROM: Projection-based nonlinear reduced order modeling with smooth neural fields.Jour- nal of Computational Physics, 532:113957, 2025

    work page 2025

Show all 48 references
  1. [9]

    Aviral Prakash and Yongjie J. Zhang. Data-driven identification of stable sparse differ- ential operators using constrained regression.Computer Methods in Applied Mechanics and Engineering, 429:117149, 2024

    2024

  2. [10]

    Aviral Prakash and Yongjie J. Zhang. Nonintrusive projection-based reduced order modeling using stable learned differential operators.Computer Methods in Applied Mechanics and Engineering, 442:117946, 2025

    2025

  3. [11]

    Brunton, Joshua L

    Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equationsfromdatabysparseidentificationofnonlinear dynamicalsystems.Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016

    2016

  4. [12]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. InProceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS’18, page 6572–6583, 2018

    2018

  5. [13]

    Kookjin Lee and Eric J. Parish. Parameterized neural ordinary differential equations: Applications to computational physics problems.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 477, 2021

    2021

  6. [14]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InProceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 6000–6010, Red Ho...

    2017

  7. [15]

    Transformers as meta-learners for implicit neural representations

    Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. InEuropean Conference on Computer Vision, pages 170–187, 2022. 38

    2022

  8. [16]

    Fries, Xiaolong He, and Youngsoo Choi

    William D. Fries, Xiaolong He, and Youngsoo Choi. LaSDI: Parametric latent space dynamics identification.Computer Methods in Applied Mechanics and Engineering, 399:115436, 2022

    2022

  9. [17]

    Park, William Fries, Daniel A

    Christophe Bonneville, Xiaolong He, April Tran, Jun S. Park, William Fries, Daniel A. Messenger, Siu W. Cheung, Yeonjong Shin, David M. Bortz, Debojyoti Ghosh, Jiun- Shyan Chen, Jonathan Belof, and Youngsoo Choi. A comprehensive review of latent space dynamics identification a...

    2024 arXiv

  10. [18]

    Messenger and David M

    Daniel A. Messenger and David M. Bortz. Weak SINDy: Galerkin-based data-driven model selection.Multiscale Modeling & Simulation, 19(3):1474–1497, 2021

    2021

  11. [19]

    Continuous PDE dynamics forecasting with implicit neural represen- tations.arXiv:2209.14855, 2023

    Yuan Yin, Matthieu Kirchmeyer, Jean-Yves Franceschi, Alain Rakotomamonjy, and Patrick Gallinari. Continuous PDE dynamics forecasting with implicit neural represen- tations.arXiv:2209.14855, 2023

    2023 arXiv

  12. [20]

    Reduced-order modeling for parame- terized PDEs via implicit neural representations.arXiv:2311.16410, 11 2023

    Tianshu Wen, Kookjin Lee, and Youngsoo Choi. Reduced-order modeling for parame- terized PDEs via implicit neural representations.arXiv:2311.16410, 11 2023

    2023 arXiv

  13. [21]

    Karniadakis

    Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, 2019

    2019

  14. [22]

    Conservative model reduction for finite-volume models.Journal of Computational Physics, 371:280–314, 2018

    Kevin Carlberg, Youngsoo Choi, and Syuzanna Sargsyan. Conservative model reduction for finite-volume models.Journal of Computational Physics, 371:280–314, 2018

    2018

  15. [23]

    Carlberg

    Kookjin Lee and Kevin T. Carlberg. Deep conservation: A latent-dynamics model for exact satisfaction of physical conservation laws.Proceedings of the AAAI Conference on Artificial Intelligence, 35(1):277–285, May 2021

    2021

  16. [24]

    Mahoney, and Aditi S

    Geoffrey Negiar, Michael W. Mahoney, and Aditi S. Krishnapriyan. Learning differen- tiable solvers for systems with hard constraints. In11th International Conference on Learning Representations, ICLR 2023, 2023

    2023

  17. [25]

    Towards enforcing hard physics constraints in operator learning frameworks

    Valentin Duruisseaux, Miguel Liu-Schiaffini, Julius Berner, and Anima Anandkumar. Towards enforcing hard physics constraints in operator learning frameworks. InICML 2024 AI for Science Workshop, 2024

    2024

  18. [26]

    Eike H. Müller. Exact conservation laws for neural network integrators of dynamical systems.Journal of Computational Physics, 488:112234, 2023

    2023

  19. [27]

    Mohan, Nicholas Lubbers, Misha Chertkov, and Daniel Livescu

    Arvind T. Mohan, Nicholas Lubbers, Misha Chertkov, and Daniel Livescu. Embed- ding hard physical constraints in neural network coarse-graining of three-dimensional turbulence.Physical Review Fluids, 8, 1 2023. 39

    2023

  20. [28]

    Jack Richter-Powell, Yaron Lipman, and Ricky T. Q. Chen. Neural conservation laws: A divergence-free perspective. InAdvances in Neural Information Processing Systems, 2022

    2022

  21. [29]

    Representation of divergence-free vector fields.Quarterly of applied mathematics, 69(2):309–316, 2011

    Cristian Barbarosie. Representation of divergence-free vector fields.Quarterly of applied mathematics, 69(2):309–316, 2011

    2011

  22. [30]

    Streamfunctionsfordivergence-freevectorfields.Quarterly of applied mathematics, 79(1):163–174, 2021

    JamesP.Kelliher. Streamfunctionsfordivergence-freevectorfields.Quarterly of applied mathematics, 79(1):163–174, 2021

    2021

  23. [31]

    Nonlinear dimen- sionality reduction for parametric problems: A kernel proper orthogonal decomposition

    Pedro Díez, Alba Muixí, Sergio Zlotnik, and Alberto García-González. Nonlinear dimen- sionality reduction for parametric problems: A kernel proper orthogonal decomposition. International Journal for Numerical Methods in Engineering, 122(24):7306–7327, 2021

    2021

  24. [32]

    Operator inference for non-intrusive model reduction with quadratic manifolds.Computer Methods in Applied Mechanics and Engineering, 403:115717, 2023

    Rudy Geelen, Stephen Wright, and Karen Willcox. Operator inference for non-intrusive model reduction with quadratic manifolds.Computer Methods in Applied Mechanics and Engineering, 403:115717, 2023

    2023

  25. [33]

    Quadratic approximation manifold for mitigating the Kolmogorov barrier in nonlinear projection-based model order reduction.Journal of Computational Physics, 464:111348, 2022

    Joshua Barnett and Charbel Farhat. Quadratic approximation manifold for mitigating the Kolmogorov barrier in nonlinear projection-based model order reduction.Journal of Computational Physics, 464:111348, 2022

    2022

  26. [34]

    Carlberg

    Kookjin Lee and Kevin T. Carlberg. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404:108973, 2020

    2020

  27. [35]

    A fast and ac- curate physics-informed neural network reduced order model with shallow masked au- toencoder.Journal of Computational Physics, 451:110841, 2022

    Youngkyu Kim, Youngsoo Choi, David Widemann, and Tarek Zohdi. A fast and ac- curate physics-informed neural network reduced order model with shallow masked au- toencoder.Journal of Computational Physics, 451:110841, 2022

    2022

  28. [36]

    Learning implicit fields for generative shape modeling

    Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5932–5941, 2019

    2019

  29. [37]

    DeepSDF: Learning continuous signed distance functions for shape representation

    JeongJ.Park, PeterFlorence, JulianStraub, RichardNewcombe, andStevenLovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    2019

  30. [38]

    Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. InProceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, 2020

    2020

  31. [39]

    Chen, Jinxu Xiang, Dong H

    Peter Y. Chen, Jinxu Xiang, Dong H. Cho, Yue Chang, G. A. Pershing, Henrique T. Maia, Maurizio M. Chiaramonte, Kevin T. Carlberg, and Eitan Grinspun. CROM: Continuous reduced-order modeling of PDEs using implicit neural representations. In The Eleventh International Conference...

    2023

  32. [40]

    Brunton, and José N

    Shaowu Pan, Steven L. Brunton, and José N. Kutz. Neural implicit flow: A mesh- agnostic dimensionality reduction paradigm of spatio-temporal data.Journal of Ma- chine Learning Research, 24(41):1–60, 2023

    2023

  33. [41]

    Learning the intrinsic dynamics of spatio-temporal processes through latent dynamics networks.Nature Communications, 15, 12 2024

    Francesco Regazzoni, Stefano Pagani, Matteo Salvador, Luca Dedé, and Alfio Quar- teroni. Learning the intrinsic dynamics of spatio-temporal processes through latent dynamics networks.Nature Communications, 15, 12 2024

    2024

  34. [42]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2017

    2017 arXiv

  35. [43]

    Peter J. Schmid. Dynamic mode decomposition of numerical and experimental data. Journal of Fluid Mechanics, 656:5–28, 2010

    2010

  36. [44]

    LeVeque.Finite Volume Methods for Hyperbolic Problems

    Randall J. LeVeque.Finite Volume Methods for Hyperbolic Problems. Cambridge Texts in Applied Mathematics. Cambridge University Press, 2002

    2002

  37. [45]

    Mixed finite elements inR3.Numerische Mathematik, 35:315–341, 1980

    Jean-Claude Nedelec. Mixed finite elements inR3.Numerische Mathematik, 35:315–341, 1980

    1980

  38. [46]

    Karakashian and Wadi N

    Ohannes A. Karakashian and Wadi N. Jureidini. A nonconforming finite element method for the stationary Navier-Stokes equations.SIAM journal on numerical analy- sis, 35(1):93–120, 1998

    1998

  39. [47]

    Efficient implementation of weighted ENO schemes.Journal of Computational Physics, 126(1):202–228, 1996

    Guang-Shan Jiang and Chi-Wang Shu. Efficient implementation of weighted ENO schemes.Journal of Computational Physics, 126(1):202–228, 1996

    1996

  40. [48]

    Efficientimplementationofessentiallynon-oscillatory shock-capturing schemes.Journal of Computational Physics, 77(2):439–471, 1988

    Chi-WangShuandStanleyOsher. Efficientimplementationofessentiallynon-oscillatory shock-capturing schemes.Journal of Computational Physics, 77(2):439–471, 1988. 41

    1988

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.